Post training 4-bit quantization of convolutional networks for rapid-deployment
Authors: Ron Banner, Yury Nahshan, Daniel Soudry
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Combining these methods, our approach achieves accuracy that is just a few percents less the state-of-the-art baseline across a wide range of convolutional models. The source code to replicate all experiments is available on Git Hub: https://github.com/submission2019/cnn-quantization. This section reports experiments on post-training quantization using six convolutional models originally pre-trained on the Image Net dataset. |
| Researcher Affiliation | Collaboration | Intel Artificial Intelligence Products Group (AIPG)1 Technion Israel Institute of Technology2 |
| Pseudocode | No | The paper contains mathematical formulations and derivations, but no explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code to replicate all experiments is available on Git Hub: https://github.com/submission2019/cnn-quantization. |
| Open Datasets | Yes | This section reports experiments on post-training quantization using six convolutional models originally pre-trained on the Image Net dataset. |
| Dataset Splits | Yes | Table 1: Image Net Top-1 validation accuracy with post-training quantization using the three methods suggested by this work. |
| Hardware Specification | No | The paper does not specify the hardware used for its experiments. |
| Software Dependencies | No | The paper mentions GEMMLOWP but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | This section reports experiments on post-training quantization using six convolutional models originally pre-trained on the Image Net dataset. We consider the following baseline setup: Per-channel-quantization of weights and activations: ... Fused Re LU: ... We use the common practice to quantize the first and the last layer as well as average/max-pooling layers to 8-bit precision. |